US ICER ASSESSMENT APPROACHES TO SURVIVAL EXTRAPOLATION: A REVIEW AND CASE STUDY
Author(s)
Tingting Qu, PhD1, Marko Zivkovic, PhD1, Zhengfan Wang, PhD2, Agota Szende, PhD3;
1Genesis Research Group, Hoboken, NJ, USA, 2Genesis Research Group, Boston, MA, USA, 3Genesis Research Group, London, United Kingdom
1Genesis Research Group, Hoboken, NJ, USA, 2Genesis Research Group, Boston, MA, USA, 3Genesis Research Group, London, United Kingdom
OBJECTIVES: Survival extrapolation is a major driver of cost-effectiveness model outcomes in oncology evaluations. The Institute for Clinical and Economic Review (ICER)’s survival curve extrapolation and selection and model transparency vary across assessments, creating uncertainty for those relying on the estimates. We reviewed ICER’s extrapolation practices and assessed how broader model consideration may improve fitness and robustness.
METHODS: All ICER assessment evidence reports published between 2016 and 2025 (n=88) were reviewed; seven (8%) conducted survival extrapolation which consistently applied the same five candidate models (exponential, log-normal, log-logistic, Gompertz, Weibull) and selected base case curves primarily using Akaike Information Criterion (AIC). The most recent case, the 2021 multiple myeloma assessment on idecabtagene vicleucel (ide-cel), was examined in depth. Its overall survival (OS) Kaplan-Meier (KM) curve was digitized, individual patient data were reconstructed using Guyot et al. (2012) algorithm, and an expanded set of candidate models (gamma, generalized gamma, and spline models) was compared in terms of fitness (AIC/Bayesian Information Criterion [BIC]) and robustness.
RESULTS: For extrapolating ide-cel's OS, ICER selected a log-normal curve. In our replicated analyses, a gamma curve demonstrated better fit (lowest AIC/BIC) than all ICER-tested models. Furthermore, a 1-knot hazard spline model produced nearly identical OS estimates to gamma (±0.1% through Month 24), supporting model robustness. In contrast, the log-normal curve overestimated OS beyond Month 20, resulting in higher undiscounted lifetime life years (LYs) (3.3) compared with the gamma and the spline model (2.3, and 2.2, respectively).
CONCLUSIONS: Our findings suggest that ICER’s restricted set of survival candidate models may overlook better-fitting alternatives. Incorporating additional model candidates like spline models, routinely reporting both AIC and BIC, and more transparently documenting visual inspection and clinical plausibility validation would enhance reproducibility and reduce uncertainty in LY estimates - ultimately improving the accuracy of cost-effectiveness results.
METHODS: All ICER assessment evidence reports published between 2016 and 2025 (n=88) were reviewed; seven (8%) conducted survival extrapolation which consistently applied the same five candidate models (exponential, log-normal, log-logistic, Gompertz, Weibull) and selected base case curves primarily using Akaike Information Criterion (AIC). The most recent case, the 2021 multiple myeloma assessment on idecabtagene vicleucel (ide-cel), was examined in depth. Its overall survival (OS) Kaplan-Meier (KM) curve was digitized, individual patient data were reconstructed using Guyot et al. (2012) algorithm, and an expanded set of candidate models (gamma, generalized gamma, and spline models) was compared in terms of fitness (AIC/Bayesian Information Criterion [BIC]) and robustness.
RESULTS: For extrapolating ide-cel's OS, ICER selected a log-normal curve. In our replicated analyses, a gamma curve demonstrated better fit (lowest AIC/BIC) than all ICER-tested models. Furthermore, a 1-knot hazard spline model produced nearly identical OS estimates to gamma (±0.1% through Month 24), supporting model robustness. In contrast, the log-normal curve overestimated OS beyond Month 20, resulting in higher undiscounted lifetime life years (LYs) (3.3) compared with the gamma and the spline model (2.3, and 2.2, respectively).
CONCLUSIONS: Our findings suggest that ICER’s restricted set of survival candidate models may overlook better-fitting alternatives. Incorporating additional model candidates like spline models, routinely reporting both AIC and BIC, and more transparently documenting visual inspection and clinical plausibility validation would enhance reproducibility and reduce uncertainty in LY estimates - ultimately improving the accuracy of cost-effectiveness results.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR11
Topic
Methodological & Statistical Research
Disease
SDC: Oncology